Research On Medical Image Segmentation Algorithm Based On UNet Model

Medical image segmentation is a very important part of medical image processing,Its development will have an important impact on visualization technology,disease diagnosis and postoperative evaluation.The location and size of lesions or organs in slice images are different,and in medical images,there are common effects such as contour blur and noise interference.Hence,how to carry out real-time image segmentation and ensure accurate effect has great practical value and theoretical research significance for the in-depth research of medical image segmentation technology.With the development of deep learning,convolutional neural network is also used by more researchers in the field of medical image segmentation.The proposal of UNet also made itself a mainstay in this field.The design of jump connection and symmetrical encode-decode structure in UNet can extract the high-level semantic information and low-level features of medical images.And its lightweight advantage can make it better adapt to smaller medical image data.In addition,the segmentation effect of local details is outstanding.Therefore,various improved networks based on UNet are also emerging in endlessly.However,continuous convolution and pooling will lead to the loss of some spatial information.Thus,this thesis proposes two improved UNet models.The first proposed network combines the context extraction structure with the long-term dependence of Trans UNet capture sequence to supplement more detailed feature information.The context extraction structure mainly includes two blocks:dense atrous convolution（DAC）module and residual multi-kernel pooling（RMP）module.The second proposed network combines the context extraction module on the basis of attention UNet and the fusion idea of the first proposed network.In order to combine different information and context modules,we also propose a transformation fusion layer to combine the high-level semantic information and multi-scale features captured by the context extraction module.In this thesis,the two proposed models are applied to eyeball vessel segmentation task and lung segmentation task.In order to ensure the advantages of each comparison model,we combines many training strategies and skills.The experimental results show that the first proposed network is superior to Trans UNet in two 2D medical image segmentation tasks.The improvement effect is better on the lung segmentation dataset,and its Dice coefficient is improved by 1.01%.Compared with all other models,the second proposed network achieves the best results in two medical segmentation tasks,such as IOU and Dice coefficients.On the lung segmentation dataset with more obvious improvement effect,the Dice coefficient of the second proposed network is improved by 1.56%compared with the 2^ndranked network.The experimental results show that the improved model proposed in this paper is better than the original UNET and other improved networks in two medical image segmentation tasks,which directly proves the effectiveness of the proposed improved model in the practical application of medical image segmentation.